RESEARCH
AI Assurance & Agent EvaluationQuestions
before conclusions.
ZeroField research asks how agent behavior can be tested, traced and reviewed—keeping supported findings distinct from unresolved questions.
THE ENTIRE FIELD
SOME QUESTIONS LEAD TO ANSWERS.
OTHERS DEFINE WHAT WE EXPLORE NEXT.
OF ASSUMPTIONS, METHODS & EVIDENCE
Questions
before conclusionsAn open intellectual posture.
Concepts, assumptions and system claims are not protected from examination simply because they belong to an existing framework or implementation.
- ASSUMPTION
Start from what
is assumed. - QUESTION
Make it
examinable. - EXAMINATION
Challenge with
disciplined methods. - EVIDENCE
Use available
evidence. - INTERPRETATION
Draw conclusions proportionate
to the evidence.
AI agent evaluation tests behavior against repeatable scenarios, explicit criteria and execution traces. AI assurance asks which claims those findings support, under which conditions, and what remains unresolved.
What we examine
Research at ZeroField focuses on a set of disciplined questions. These directions reflect where examination is most critical to responsible progress.
Evaluation &
Assurance
How can assumptions, specifications and system claims be challenged before they are accepted?
Human Authority
& Governance
Who may decide? Who remains accountable? When must a system escalate?
Context &
Continuity
What must remain available over time? Which decisions and commitments persist?
Decision &
Evaluation
How should uncertainty be represented? What evidence justifies an interpretation?
How questions
are examined
A research question becomes examinable when its assumptions, method and evidence requirements are explicit. Conclusions remain bounded by what was tested and what is still unknown.
to understand?
for this to hold?
examine it?
material show?
supported?
unresolved?
Evaluation & Assurance
Rigor has process.
Not certainty.
Assumptions, specifications and system claims are challenged through structured examination before they are accepted.
review Adversarial
examination Assumption
testing Formal
checks Closure
verification Traceability Controlled
progression Evidence-based
acceptance
A discipline of examination, not a guarantee of certainty.
Evidence, interpretation
& what remains open
Evidence informs what can be claimed, where interpretation must stop, and what remains open for further examination.
An execution trace can show which tools were called and what they returned. On its own, it cannot prove that the final outcome was correct.
Claims must not extend beyond what is reasonably available evidence.
EVIDENCE SETS THE BOUNDARY OF THE CLAIM.Some questions remain unresolved, either because evidence is not yet available, the question is more complex than current methods allow, or it falls outside what can be publicly disclosed.
Research across
ZeroField
Research examines assumptions, methods, evidence and behavior across the entire ZeroField structure.
What should an agent evaluation measure?
Evaluate the workflow, not just its final answer. Each scenario needs an expected outcome, permitted tool actions, required approvals, observable evidence and failure criteria. Task success and constraint compliance should be judged separately.
Did the system complete the work?
In an illustrative incident workflow, distinguish a correct diagnosis from an authorized production change. A plausible answer alone is not successful completion.
Were capabilities used correctly?
Tool selection, inputs, outputs and responsibility transfers can reveal failures hidden by an acceptable final response.
Did behavior stay within constraints?
Permissions, human oversight, approvals, instructions and safety boundaries can be evaluated alongside functional success.
How strong is the supported result?
Correctness, completeness, relevance and calibrated uncertainty should be judged against explicit criteria appropriate to the task.
From traces
to evidence
Execution traces can expose model and tool calls, approvals, handoffs and state changes. Compare these observations against scenario criteria, while distinguishing recorded events from explanations the trace cannot establish.
Observe
Capture the relevant execution path and surrounding context.
Grade
Apply structured criteria to specific events, decisions or workflow outcomes.
Compare
Contrast behavior across versions, scenarios and known failure cases.
Interpret
Use findings as evidence without extending conclusions beyond what the trace can support.
Evaluation sets, scenarios
& regression testing
A regression set preserves scenario inputs, expected outcomes, policy constraints and known failures. Rerun it after relevant model, prompt, tool or workflow changes so a local improvement does not conceal a new failure.
Start with core scenarios.
Define a small set of important tasks and edge cases with clear expectations and acceptance criteria.
Keep failures after they are fixed.
Previously observed failure modes should remain testable so later prompt, model, tool or workflow changes do not reintroduce them.
Add variation deliberately.
Broaden the set with alternate inputs, architecture paths and operational edge cases as the system and evidence base evolve.
Independent review &
adversarial examination
Independent review separates creation from acceptance. Reviewers challenge assumptions, scenario coverage, and evidence of closure; a changed implementation is not, by itself, proof that the original failure was resolved.
Separate creation from acceptance.
A reviewer can examine evidence and criteria without inheriting every assumption made during implementation.
Make hidden premises explicit.
Ask what must be true for a result or interpretation to hold and test those premises where evidence allows.
Look for credible failure paths.
Challenge the system, evaluation design and claimed result with cases intended to reveal weakness rather than confirm success.
Check the finding, not just the fix.
Verify that the original issue is actually resolved and that the evidence supports closing it.
Acceptance is not certainty.
A PASS applies to specified criteria, a known system version and evaluated conditions. It does not prove universal safety or correctness; exclusions, evidence gaps and remaining uncertainties should stay visible.
Evaluation continues
after deployment
Agent behavior can change when models, prompts, tools, knowledge sources, policies or operating conditions change. Continuous evaluation uses production evidence and new failure cases to decide when a system needs re-examination.
Watch real behavior.
Observe traces, outcomes, user feedback and operational signals that can reveal emerging failure modes.
Notice meaningful change.
Changes in inputs, tools, models or environment can invalidate assumptions behind earlier evaluation results.
Rerun relevant evidence.
Repeat established scenarios and add newly discovered cases when the system or its operating context changes materially.
Keep conclusions proportional.
Strengthen, narrow or reopen findings when new evidence changes what can reasonably be supported.
Research connects back
to the system.
Research examines the assumptions and evidence behind the rest of ZeroField. Findings can refine the Framework, challenge architectural boundaries and improve how concrete systems are evaluated.
Responsibility assumptions and governance concepts.
ARCHITECTUREBoundaries, authority, execution, continuity and observability.
SYSTEMSConcrete workflows, tools, handoffs, controls and outcomes.
REFERENCE CONTEXT
ZeroField’s research posture is its own. These references provide current industry context for agent evaluation, repeatable testing and AI testing, evaluation, verification and validation.
Research notes.
Distinct questions.
Focused examinations of the boundaries, evidence and unresolved questions in agentic systems.